lauvshree / asladigitechlearningmaterial Goto Github PK

Looks like an overall good analysis @lauvshree! I put my thoughts below. Use what you choose :)

• Can you clarify what "previous data" means in "purchasing term deposits based on the previous data"? This means historical data, right? Also it's not just historical data, it also includes demographic data. Maybe change historical data to "based on the customer profile and banking behaviours"?
• Why is the "day" column being checked as a non-numeric?
• Just to clarify, the previous marketing campaign is a different campaign for the current campaign on term deposits?
• I don't know how much documentation you need on this, and how pretty it should be, but the area starting with looking at "maximum days lapsed since contact" could use some markdown prior to the code cells if you do need this to be well documented.
• Chi-squared is tricky to use because it is highly sensitive (rejects easily). Do you think, based upon the fact that every chi-squared was highly significant, that the red/blue bar graphs are worth putting down?
• Also, I would maybe move the sentence "From the description of the dataset done above, it is evident that while some of the features which are continuously distributed are normal, some are not." to after the pairplots.
• Why did you choose the ordered numerics you did for "Employment", "MartialStatus", and "PreviousOutcome"? I am not necessarily saying what you chose is wrong, but why does "entrepreneur" deserve to be the number "11", while "self-employed" deserves the smaller number "9"? Does this ordering affect the outcome?
• Why did you choose to do the gaussian naive bayes first? Should you put it's own sub-heading over this model, like you do for the ensemble classifiers?
• Also why balance after running NB? Shouldn't you run balance before-hand?
• Could you show the max_depth used in the non-regularised DecisionTreeClassifier? Maybe you don't need this...but I'd be curious how much it changed. Also your non-regularised model still performed very well on the test set, making me think that a max_depth > 5 is still the right way to go. K-fold cross-val would prove this out (if the higher depth performs well on many different test sets)
• See the above go through similarly when you use the RandomForestClassifier without a max_depth. If you set your max_depth to 5 on the RF.
• I always thought ensemble methods designed a bunch of different classifiers, and then predicted the test set values using all of the classifiers, and then chose the majority prediction from the collection of classifiers over just one. Should there be a section here where you aggregate all the classifiers together, and then do some final test set accuracy?